Tuesday, June 21, 2011

Glicko

I said earlier that I wouldn't be assessing the Glicko rating system because it was basically the same as TrueSkill, so I didn't expect significantly better performance than TrueSkill.  That hasn't changed, but I decided to take a look at Glicko anyway, mainly because I've become intrigued with the Whole History Rating algorithm from RĂ©mi Coulom and that led (indirectly) back to Glicko. I haven't yet wrapped my head around WHR (and may never do so -- the math is difficult) but I was interested enough to code up the Glicko algorithm.

I won't attempt to explain the Glicko algorithm -- it's described in detail on Mark Glickman's website -- but just jump into the performance.  My implementation recalculates a team's rating after every game -- the same as done with TrueSkill.  There is very little to tweak in Glicko, so here's the basic performance:

  Predictor    % Correct    MOV Error  
TrueSkill72.8%11.09
Glicko70.8%11.71

As with TrueSkill we can hack an MOV-based factor into Glicko and improve our accuracy somewhat:

  Predictor    % Correct    MOV Error  
TrueSkill72.8%11.09
Glicko70.8%11.71
Glicko + MOV71.8%11.56

Even so Glicko isn't a competitor for "best of" the rating systems.

No comments:

Post a Comment